Teach Kaun's text reader about encodings

[tmattio]

Why it matters

Training logs often contain accented characters or non-English text. Right now Kaun.Dataset.from_text_file ignores its encoding argument and just returns raw bytes. As soon as you point it at UTF-8 or Latin-1 files, you risk broken strings or exceptions, which makes the monitoring dashboard unusable on real datasets.

How to see the gap

Skim kaun/lib/kaun/dataset/dataset.ml, around from_text_file. The function stores encoding in _ and never decodes the file. If you create a small UTF-8 file with emoji and iterate over the dataset, the text comes back as mangled characters.

Your task

• Honor the encoding parameter in from_text_file (and the helpers that call it) by decoding each chunk before splitting on newlines.
• Add tests in kaun/test/test_dataset.ml that cover UTF-8 and Latin-1 snippets so we know the decoding works.
• Make sure the default behaviour stays the same when callers do not pass ~encoding.

Tips

• The Uutf library is already available through Raven; it can decode incrementally from a Bigarray-backed string.
• Keep the chunked reading logic intact—just convert the bytes to OCaml strings with the right encoding as they arrive.
• Use Filename.temp_file (already in the test helpers) to build short fixtures that contain characters outside plain ASCII (for example, describe an emoji with its code point).

Done when

• Passing a non-default ~encoding produces correctly decoded strings.
• The dataset tests cover at least one UTF-8 example and one Latin-1 example.
• dune runtest kaun passes after your changes.

[Satarupa22-SD]

@tmattio Please assign this issue to me. Thanks!

[tmattio]

@Satarupa22-SD sure, go ahead!